Image: Iryna Imago / Shutterstock.com

A man from Florida won’t confide his darkest thoughts to anyone but his ex-girlfriend—but he does share them with a chatbot. For months, he writes to ChatGPT about how he wants to win her back—and, if necessary, shoot her. What he doesn’t realize is that OpenAI is reading along. And eventually, the company itself hits the alarm button. This case shows that, legally speaking, a chat with an AI is far less private today than many users believe.

From Flowers and Reconciliation to a Murder Plot

Darren Shida Zhou, a 25-year-old financial analyst from West Palm Beach who worked at Goldman Sachs, turned to ChatGPT after a six-month relationship ended. At first, his messages were about heartbreak, jealousy, and the desire to reconcile. Over time, however, the messages became more specific: He described a detailed plan to ambush his ex-partner with flowers in the parking lot of her gym. According to investigative documents, if she rejected him, he intended to pull out a gun, shoot her, and then kill himself. Later, he reportedly even spoke to the chatbot about kidnapping and violence against her family members.

At the same time, according to investigators, he bombarded the young woman with calls and messages—some of them from different phone numbers he obtained through apps after she had blocked him. She described him to the police as unpredictable, jealous, and controlling.

The Moment When AI Becomes a Witness

This is exactly where the story gets legally interesting. OpenAI uses automated systems that scan chat histories for signs of violence or rule violations. If messages raise red flags, trained staff examine the context more closely. In the vast majority of cases, this results in an account suspension—end of story, next user.

But if OpenAI detects an “immediate and credible risk” to a person, the case goes a step further: to law enforcement. That is exactly what happened here. The FBI received about two months’ worth of chat logs—note that these were only the man’s own messages, not the AI’s responses—and forwarded them to investigators in Palm Beach County. A review by the police concluded that this was not merely vague angry ranting following a breakup, but rather a recurring pattern of planning and preparation. The officers also cross-referenced the statements with messages the man had sent directly to his ex-girlfriend.

Six days later, in late May, the man was arrested. He spent two days in pretrial detention, then was released on $100,000 bail. His employer, Goldman Sachs, terminated his employment in June.

A Lenient Sentence Despite Serious Allegations

The prosecution charged him with aggravated stalking, written death threats, and the misuse of communication devices—crimes that, in theory, carry long prison sentences in the United States. On August 13, he pleaded guilty to all charges. The presiding judge nevertheless sentenced him to “only” eight years of probation, with electronic monitoring for the first two years, including a no-contact order and a weapons ban. A formal conviction was suspended—a procedure common in Florida to which his ex-partner had agreed. This spares him, at least formally, a criminal record.

When the Machine Becomes a Key Witness

Anyone who thinks chatting with an AI is like keeping a diary under your pillow is sorely mistaken. For us, this case serves as a wake-up call in two ways. First: An AI that actively involves law enforcement agencies is no longer a “gray-area” technology, but is in fact a participant in criminal proceedings—without any judicial oversight, without any traditional rules of evidence, based solely on internal company criteria. That can save lives—no question about it. But it also raises the uncomfortable question of who actually determines when a company like OpenAI considers the threshold for a “credible risk” to have been crossed—and when it has not. This is precisely where the company demonstrably failed in another, more tragic case in Canada.

Second, and this is, in our view, the real scandal hidden in the fine print: The ease with which people today entrust their most dangerous thoughts to a machine is fundamentally changing the evidence in criminal proceedings. What used to have to be painstakingly reconstructed is now neatly logged and time-stamped. Our prediction: In the coming years, chat histories with AI systems will become just as commonplace as evidence as WhatsApp messages or emails are today.